Speech Emotion Recognition In Continuous Space Using IEMOCAP Database

Abstract

The exploration of emotion recognition from speech signals has garnered considerable attention across various domains. Instead of categorizing emotions into discrete classes such as anger and happiness, the dimensional continuous model is quite common. This study focuses on predicting arousal, valence, and dominance. Constructing a dimensional speech emotion recognition system involves selecting appropriate feature extraction and classification methods, a topic that remains under discussion. To tackle this challenge, this paper introduces a promising approach for extracting High Statistical Functions (HSFs) features from speech and evaluates two deep learning models for predicting arousal, valence, and dominance. The proposed models provide insights into the potential of optimized convolutional neural networks (CNN1D) and Long Short-Term Memory (LSTM) algorithms for emotion recognition tasks in speech signals. The optimized algorithms showed a significant result across both speaker-dependent and speaker-independent scenarios when evaluated on the IEMOCAP dataset.